skillstat

English | 中文

Audit skills across Claude Code, Codex, and Cursor. You've installed dozens of agent skills. Which ones actually fire, in which agent, and which are just sitting on disk? skillstat reads local transcript evidence and tells you.

• 🔍 Cross-agent evidence — one view for Claude Code, Codex, and Cursor, including per-agent counts, last use, projects, and trigger source.
• 💸 Claude context cost — estimates how many tokens Claude's skill listing injects per session and how much is spent on skills you never use.
• ✂️ Reversible slimming — moves idle skills into skills-disabled/ so you can restore any of them with a single mv.
• 🔒 Local & AI-free — pure deterministic parsing. Nothing leaves your machine, no model calls, no tokens burned to run it.

$ skillstat cost
skillstat cost  (skill_listing context overhead, estimated)

METRIC                          VALUE
Skills offered per session        121
Installed skills (disk)           128
Idle offered skills (>30d)        116
Skill-listing tokens / session  ~3.9k
  ↳ wasted on idle skills       ~3.7k
Sessions analyzed                  38

116 of 121 offered skills have been idle ≥30d, costing ~3.7k tokens per session.

Install

npm install -g skillstat
# or run without installing:
npx skillstat

Requires Node ≥ 18. No other dependencies.

Use it as an agent skill

skillstat ships a SKILL.md, so your coding agent can run it for you — just ask "which of my skills do I actually use?" or "what's filling my context?" and it'll pick the right sub-command.

# Claude Code / Codex / Cursor (Agent Skills compatible):
npx skills add no7z/skillstat

Or drop it in manually — copy this repo's SKILL.md into ~/.claude/skills/skillstat/SKILL.md (Claude Code) or ~/.codex/skills/… (Codex). Then invoke it explicitly with /skillstat, or let it auto-activate from the description. The skill just shells out to the skillstat CLI, so keep the CLI installed (or it falls back to npx).

Usage

skillstat                    # scan: per-skill trigger counts (default)
skillstat scan --all         # include offered-but-never-triggered skills
skillstat scan --agent codex # one source; comma-separated combinations work
skillstat cost               # how much context are idle skills costing?
skillstat report -o r.html   # self-contained HTML report (offline, shareable)
skillstat slim --days 60     # archive skills idle 60+ days (asks first)
skillstat slim --restore     # undo: move archived skills back

Commands

Command	What it does
scan	Table of every skill: fires, agent, evidence type, last-fired, projects.
cost	Claude-only skill_listing token overhead per session and the share wasted on idle skills.
report	Writes a self-contained dark-themed HTML dashboard (double-click, no server).
slim	Moves idle user skills (never plugins) to ~/.claude/skills-disabled/. Reversible (--restore), confirms first.

Options

Flag	Meaning
-d, --days <n>	Idle threshold for "zombie" skills (default 30).
--agent <list>	claude, codex, cursor, or a comma-separated combination (default all). --source is an alias.
-a, --all	scan: also list offered skills that never triggered.
-o, --out <f>	report: output path (default skillstat-report.html).
-y, --yes	slim: skip the confirmation prompt.
--restore	slim: move everything in skills-disabled/ back into skills/.
--json	Machine-readable output for scan and cost.

How it works

The agents expose different evidence, so skillstat reports only what each local format can prove:

Agent	Transcript evidence	Installed-skill roots
Claude Code	Skill tool calls (explicit) and invoked_skills (auto)	~/.claude/skills, ~/.claude/plugins
Codex	explicit /skill user messages and tool calls that read a skill's SKILL.md (observed)	~/.codex/skills, ~/.agents/skills, Codex plugin cache
Cursor	explicit /skill user messages	~/.cursor/skills

Cursor transcript events do not currently contain timestamps, so skillstat uses the transcript file's modification time for last-fired. Cursor auto-activation and Codex offered-list/context cost are not claimed because those signals are not present in their local transcript formats.

For context cost it reads the skill_listing attachment injected into each session and estimates its token size (heuristic, no tokenizer — kept dependency-free and offline). cost is therefore Claude-only, even in an all-agent scan.

Set CLAUDE_CONFIG_DIR, CODEX_HOME, or CURSOR_CONFIG_DIR to point at a non-default config location.

Caveats

• Token figures are estimates (chars/words heuristic), meant for relative comparison, not billing.
• Transcript formats are internal implementation details and may change. skillstat degrades gracefully (skips lines it can't parse) rather than crashing.
• slim only ever touches skills under ~/.claude/skills/ — plugin-provided skills and Codex/Cursor installs are read-only and never moved.

License

MIT